Francesco Petrogalli via llvm-dev
2019-Mar-14 14:27 UTC
[llvm-dev] Scalable Vector Types in IR - Next Steps?
> On Mar 14, 2019, at 5:45 AM, Graham Hunter via llvm-dev <llvm-dev at lists.llvm.org> wrote: > > We would like to be able to reach consensus on an approach around the end > of EuroLLVM this year so that we can begin a full implementation.EuroLLVM will be a very good occasion to hear all the opinions and reach consensus on the approach. @Graham, maybe you should set up a round table open to anyone, and make sure that key people (third parties asking for a different approach, people involved in this discussion) will receive the invite? Francesco
David Greene via llvm-dev
2019-Mar-15 03:02 UTC
[llvm-dev] Scalable Vector Types in IR - Next Steps?
Francesco Petrogalli via llvm-dev <llvm-dev at lists.llvm.org> writes:>> On Mar 14, 2019, at 5:45 AM, Graham Hunter via llvm-dev <llvm-dev at lists.llvm.org> wrote: >> >> We would like to be able to reach consensus on an approach around the end >> of EuroLLVM this year so that we can begin a full implementation. > > EuroLLVM will be a very good occasion to hear all the opinions and reach consensus on the approach. > > @Graham, maybe you should set up a round table open to anyone, and > make sure that key people (third parties asking for a different > approach, people involved in this discussion) will receive the invite?We tried two round tables at the Nov. LLVMDev and no serious objections were raised, but we knew we didn't have all the right people there. I am somewhat skeptical another roundtable without commitment to attend from all able parties ahead of time will accomplish much. Speaking for myself (and not Cray), it is frustrating to have had a bunch of discussion on the mailing list and in reviews where concerns were raised and to see a lot of radio silence to responses to those concerns, only to see a message about a potential change in direction driven by off-list discussions where concerns and responses to concerns are unknown and therefore not addressable. I completely understand that ARM needs to make progress and I very much want to see that progress. I just don't want to see a Plan B leading to a situation where VLA support doesn't ever make it into LLVM. It is somewhat embarrassing that gcc already has a release with VLA support for SVE and LLVM is stuck in the starting blocks. -David
Graham Hunter via llvm-dev
2019-Mar-15 10:18 UTC
[llvm-dev] Scalable Vector Types in IR - Next Steps?
Hi David,> We tried two round tables at the Nov. LLVMDev and no serious objections > were raised, but we knew we didn't have all the right people there. I > am somewhat skeptical another roundtable without commitment to attend > from all able parties ahead of time will accomplish much.Agreed, but I'll try scheduling one anyway.> Speaking for myself (and not Cray), it is frustrating to have had a > bunch of discussion on the mailing list and in reviews where concerns > were raised and to see a lot of radio silence to responses to those > concerns, only to see a message about a potential change in direction > driven by off-list discussions where concerns and responses to concerns > are unknown and therefore not addressable.I didn't want private meetings either, but repeatedly requesting public feedback for the RFC or patches hadn't provided reasoning behind any concerns that people had. The agreement reached at the meeting was for the objectors to post their reasons for objecting and counter-proposal in public so discussion could take place, and Arm would investigate the details of the counter-proposal. Unfortunately, that post never happened, so I found myself a bit stuck and had to post it for them -- not a situation I wanted. I have always wanted the discussion to take place in public.> I completely understand that ARM needs to make progress and I very much > want to see that progress. I just don't want to see a Plan B leading to > a situation where VLA support doesn't ever make it into LLVM. It is > somewhat embarrassing that gcc already has a release with VLA support > for SVE and LLVM is stuck in the starting blocks.Agreed. -Graham